The clause that blew up a $200M contract

In July 2025, the US Department of Defense awarded prototype agreements worth up to $200 million each to four AI labs — Anthropic, Google, OpenAI and xAI — to accelerate adoption of advanced AI across national-security workflows. Anthropic became the first AI lab to integrate its models into mission workflows on classified networks.

Then, in February 2026, the relationship ruptured. According to reporting compiled by TechPolicy.Press and summarised on Wikipedia, renegotiations broke down over a single contractual formulation. The Pentagon wanted language authorising Claude for 'any lawful use' — an umbrella phrase that, in Anthropic's reading, would have permitted deployment for domestic mass surveillance and for lethal targeting inside fully autonomous weapons systems operating without meaningful human authorisation.

Anthropic refused. The company wanted the military to commit, in writing, that it would not use Claude in weapons that identify and fire on targets without human input, and would not use Claude to analyse the location records, financial data and other commercially-purchased datasets of US persons.

The designation

On 27 February 2026, Defense Secretary Pete Hegseth announced the DoD would designate Anthropic a 'supply-chain risk to national security.' As Fortune and NPR reported, the designation terminates Anthropic's contract and forces entities that want to do business with the DoD to sever commercial ties with the company — with a six-month wind-down window. In parallel, President Trump ordered all federal agencies to 'immediately cease all use' of Anthropic's technology, granting the Pentagon a six-month phase-out for existing deployments.

This is an extraordinary escalation: a sitting administration directing the entire federal government to drop one named US AI company over a safety clause.

What Anthropic's policy actually says

Anthropic's usage policy prohibits using its models to develop or design weapons, and bars use of its services for lethal autonomous warfare without human oversight. Crucially, Anthropic's definition of 'autonomous weapons' is narrow: systems that select and engage targets without human supervision. That narrowness matters — and, as critics note, it leaves a large grey zone where humans remain nominally 'in the loop' (see our companion piece on the human-in-the-loop debate).

OpenAI steps in

The same day the designation landed, OpenAI announced an agreement with the Department of Defense to deploy its systems in classified environments — explicitly filling the gap Anthropic left. OpenAI said its deal also carried red lines against domestic mass surveillance and the independent direction of autonomous weapons, enforced via a cloud-only deployment architecture and cleared safety researchers 'in the loop.' After CEO Sam Altman called the initial rollout 'opportunistic and sloppy,' OpenAI amended the contract on 2 March 2026 to tighten the surveillance prohibition and exclude intelligence agencies including the NSA from that specific contract.

Why it matters

Strip away the politics and the dispute is about a genuinely hard question: can a commercial AI vendor hold a safety line against the most powerful customer it will ever have? Anthropic chose the line over the contract. The cost was its entire federal book of business. Whether that proves principled or commercially fatal is the story to watch through 2026.